06. Transfer Learning with TensorFlow Part 3: Scaling up (🍔👁 Food Vision mini)¶
In the previous two notebooks (transfer learning part 1: feature extraction and part 2: fine-tuning) we've seen the power of transfer learning.
Now we know our smaller modelling experiments are working, it's time to step things up a notch with more data.
This is a common practice in machine learning and deep learning: get a model working on a small amount of data before scaling it up to a larger amount of data.
🔑 Note: You haven't forgotten the machine learning practitioners motto have you? "Experiment, experiment, experiment."
It's time to get closer to our Food Vision project coming to life. In this notebook we're going to scale up from using 10 classes of the Food101 data to using all of the classes in the Food101 dataset.
Our goal is to beat the original Food101 paper's results with 10% of data.
Machine learning practitioners are serial experimenters. Start small, get a model working, see if your experiments work then gradually scale them up to where you want to go (we're going to be looking at scaling up throughout this notebook).
What we're going to cover¶
We're going to go through the follow with TensorFlow:
- Downloading and preparing 10% of the Food101 data (10% of training data)
- Training a feature extraction transfer learning model on 10% of the Food101 training data
- Fine-tuning our feature extraction model
- Saving and loaded our trained model
- Evaluating the performance of our Food Vision model trained on 10% of the training data
- Finding our model's most wrong predictions
- Making predictions with our Food Vision model on custom images of food
How you can use this notebook¶
You can read through the descriptions and the code (it should all run, except for the cells which error on purpose), but there's a better option.
Write all of the code yourself.
Yes. I'm serious. Create a new notebook, and rewrite each line by yourself. Investigate it, see if you can break it, why does it break?
You don't have to write the text descriptions but writing the code yourself is a great way to get hands-on experience.
Don't worry if you make mistakes, we all do. The way to get better and make less mistakes is to write more code.
📖 Resource: See the full set of course materials on GitHub: https://github.com/mrdbourke/tensorflow-deep-learning
A colourful figure of the model we've created with: 224x224 images as input, data augmentation as a layer, EfficientNetB0 as a backbone, an averaging pooling layer as well as dense layer with 10 neurons (same as number of classes we're working with) as output.
When fine-tuning and unfreezing layers of your pre-trained model, it's common practice to lower the learning rate you used for your feature extraction model. How much by? A 10x lower learning rate is usually a good place to to start.
